def main(_):
pp.pprint(flags.FLAGS.__flags)
with tf.device('/cpu:0'), tf.Session() as sess:
if FLAGS.task == 'copy':
if FLAGS.is_train:
cell, ntm = copy_train(FLAGS, sess)
else:
cell = NTMCell(input_dim=FLAGS.input_dim,
output_dim=FLAGS.output_dim,
controller_layer_size=FLAGS.controller_layer_size,
write_head_size=FLAGS.write_head_size,
read_head_size=FLAGS.read_head_size)
ntm = NTM(cell, sess, 1, FLAGS.max_length,
test_max_length=FLAGS.test_max_length, forward_only=True)
ntm.load(FLAGS.checkpoint_dir, 'copy')
copy(ntm, int(FLAGS.test_max_length*1/3), sess)
print
copy(ntm, int(FLAGS.test_max_length*2/3), sess)
print
copy(ntm, int(FLAGS.test_max_length*3/3), sess)
elif FLAGS.task == 'recall':
pass
def compareFixed():
t = Tasks()
x_test, y_test = t.sequence_type_1(100)
add_params, mul_params = torch.load('program_memory/add.pt'), torch.load(
'program_memory/mul.pt')
hnm = HNM(10, 20, add_params, mul_params)
hnm.load_state_dict(torch.load("learned_params/hnm_arch_2.pt"))
ntm = NTM(10, 20)
ntm.load_state_dict(torch.load("learned_params/ntm.pt"))
lstm = LSTM(14, 256, 325, 1)
lstm.load_state_dict(torch.load("learned_params/lstm.pt"))
hnm_diff, lstm_diff, ntm_diff = 0, 0, 0
for i in range(len(x_test)):
hnm_out = hnm.recurrent_forward(x_test[i:i + 1])
ntm_out = ntm.recurrent_forward(x_test[i:i + 1])
lstm_out = lstm.recurrent_forward(x_test[i:i + 1])
answer = np.argmax(y_test[i:i + 1].detach().numpy())
hnm_diff += abs(answer - np.argmax(hnm_out.detach().numpy()))
ntm_diff += abs(answer - np.argmax(ntm_out.detach().numpy()))
lstm_diff += abs(answer - np.argmax(lstm_out.detach().numpy()))
print(hnm_diff / len(y_test), ntm_diff / len(y_test),
lstm_diff / len(y_test))
def create_ntm(config, sess, **ntm_args):
if config.rand_hyper:
hyper_params = {}
if config.is_test:
hyper_params = load_hyperparamters(config)
else:
hyper_params = generate_hyperparams(config)
print(" [*] Hyperparameters: {}".format(hyper_params))
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=hyper_params["c_layer"],
controller_dim=hyper_params["c_dim"],
mem_size=hyper_params["mem_size"],
write_head_size=config.write_head_size,
read_head_size=config.read_head_size,
is_LSTM_mode=config.is_LSTM_mode)
scope = ntm_args.pop('scope', 'NTM-%s' % config.task)
# Description + query + plan + answer
min_length = (config.min_size -
1) + 1 + config.plan_length + (config.min_size - 1)
max_length = int(((config.max_size * (config.max_size - 1) / 2) + 1 +
config.plan_length + (config.max_size - 1)))
ntm = NTM(cell,
sess,
min_length,
max_length,
config.min_size,
config.max_size,
scope=scope,
**ntm_args,
lr=hyper_params["lr"],
momentum=hyper_params["momentum"],
decay=hyper_params["decay"],
beta=hyper_params["l2"])
else:
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
controller_dim=config.controller_dim,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size,
is_LSTM_mode=config.is_LSTM_mode)
scope = ntm_args.pop('scope', 'NTM-%s' % config.task)
# Description + query + plan + answer
min_length = (config.min_size -
1) + 1 + config.plan_length + (config.min_size - 1)
max_length = int(((config.max_size * (config.max_size - 1) / 2) + 1 +
config.plan_length + (config.max_size - 1)))
ntm = NTM(cell,
sess,
min_length,
max_length,
config.min_size,
config.max_size,
scope=scope,
**ntm_args)
return cell, ntm
def __init__(self,
d_vocab,
d_emb,
d_dec,
max_len,
bos_idx,
num_heads=8,
N=64,
M=32,
seg_size=20):
super().__init__()
self.d_vocab = d_vocab
self.seg_size = seg_size
self.embs = nn.Embedding(d_vocab, d_emb)
self.rnn = nn.GRU(d_emb, d_dec, batch_first=True)
self.ntm_scale = nn.Parameter(torch.zeros([1, d_dec]),
requires_grad=True)
self.ntm = NTM('mem-aug',
embedding_size=d_dec,
hidden_size=d_dec,
memory_size=M,
head_num=num_heads,
memory_feature_size=N,
output_size=d_dec)
self.init = nn.Parameter(torch.zeros(1, d_dec), requires_grad=True)
self.bos_idx = nn.Parameter(torch.tensor([bos_idx]),
requires_grad=False)
self.out_layer = nn.Linear(d_dec, d_vocab)
self.max_len = max_len
def trainNTM():
t = Tasks()
x_train, y_train = t.sequence_type_1(2000)
ntm = NTM(10, 20)
ntm.train(x_train, y_train, 1, maxEpoch=25, learning_rate=0.0006)
def predict_train(config, sess):
"""Train an NTM for the copy task given a TensorFlow session, which is a
connection to the C++ backend"""
if not os.path.isdir(config.checkpoint_dir):
raise Exception(" [!] Directory %s not found" % config.checkpoint_dir)
# delimiter flag-like vector inputs indicating the start and end
# you can see these in the figure examples in the README
# this is kind of defined redundantly
start_symbol = np.zeros([config.input_dim], dtype=np.float32)
start_symbol[0] = 1
end_symbol = np.zeros([config.input_dim], dtype=np.float32)
end_symbol[1] = 1
# initialise the neural turing machine and the neural-net controller thing
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size)
ntm = NTM(cell, sess, config.min_length, config.max_length*3)
print(" [*] Initialize all variables")
tf.initialize_all_variables().run()
print(" [*] Initialization finished")
start_time = time.time()
for idx in xrange(config.epoch):
# generate a sequence of random length
seq_length = randint(config.min_length, config.max_length) * 4
inc_seq, comp_seq = generate_predict_sequence(seq_length, config.input_dim - 2)
# this somehow associates the desired inputs and outputs with the NTM
feed_dict = {input_:vec for vec, input_ in zip(inc_seq, ntm.inputs)}
feed_dict.update(
{true_output:vec for vec, true_output in zip(comp_seq, ntm.true_outputs)}
)
feed_dict.update({
ntm.start_symbol: start_symbol,
ntm.end_symbol: end_symbol
})
# this runs the session and returns the current training loss and step
# I'm kind of surprised it returns the step, but whatevs
_, cost, step = sess.run([ntm.optims[seq_length],
ntm.get_loss(seq_length),
ntm.global_step], feed_dict=feed_dict)
# how does one use these checkpoints?
if idx % 100 == 0:
ntm.save(config.checkpoint_dir, 'copy', step)
if idx % print_interval == 0:
print("[%5d] %2d: %.2f (%.1fs)" \
% (idx, seq_length, cost, time.time() - start_time))
print("Training Copy task finished")
return cell, ntm
def trainNTM():
ntm = NTM(10, 14)
X, y = [], []
for i in range(10):
tempX, tempY = getData("data/observations_"+str(i*500)+".npy", "data/actions_"+str(i*500)+".npy")
X.extend(tempX)
y.extend(tempY)
print(len(X), len(y))
ntm.train(X, y, 1)
def copy_train(config):
sess = config.sess
if not os.path.isdir(config.checkpoint_dir):
raise Exception(" [!] Directory %s not found" % config.checkpoint_dir)
# delimiter flag for start and end
start_symbol = np.zeros([config.input_dim], dtype=np.float32)
start_symbol[0] = 1
end_symbol = np.zeros([config.input_dim], dtype=np.float32)
end_symbol[1] = 1
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size)
ntm = NTM(cell, sess, config.min_length, config.max_length)
print(" [*] Initialize all variables")
tf.initialize_all_variables().run()
print(" [*] Initialization finished")
start_time = time.time()
for idx in xrange(config.epoch):
seq_length = randint(config.min_length, config.max_length)
seq = generate_copy_sequence(seq_length, config.input_dim - 2)
feed_dict = {input_: vec for vec, input_ in zip(seq, ntm.inputs)}
feed_dict.update({
true_output: vec
for vec, true_output in zip(seq, ntm.true_outputs)
})
feed_dict.update({
ntm.start_symbol: start_symbol,
ntm.end_symbol: end_symbol
})
_, cost, step = sess.run([
ntm.optims[seq_length],
ntm.get_loss(seq_length), ntm.global_step
],
feed_dict=feed_dict)
if idx % 100 == 0:
ntm.save(config.checkpoint_dir, 'copy', step)
if idx % print_interval == 0:
print("[%5d] %2d: %.2f (%.1fs)" \
% (idx, seq_length, cost, time.time() - start_time))
print("Training Copy task finished")
return cell, ntm
def copy_train(config):
sess = config.sess
if not os.path.isdir(config.checkpoint_dir):
raise Exception(" [!] Directory %s not found" % config.checkpoint_dir)
# delimiter flag for start and end
start_symbol = np.zeros([config.input_dim], dtype=np.float32)
start_symbol[0] = 1
end_symbol = np.zeros([config.input_dim], dtype=np.float32)
end_symbol[1] = 1
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size)
ntm = NTM(cell, sess, config.min_length, config.max_length)
print(" [*] Initialize all variables")
tf.initialize_all_variables().run()
print(" [*] Initialization finished")
start_time = time.time()
for idx in xrange(config.epoch):
seq_length = randint(config.min_length, config.max_length)
seq = generate_copy_sequence(seq_length, config.input_dim - 2)
feed_dict = {input_:vec for vec, input_ in zip(seq, ntm.inputs)}
feed_dict.update(
{true_output:vec for vec, true_output in zip(seq, ntm.true_outputs)}
)
feed_dict.update({
ntm.start_symbol: start_symbol,
ntm.end_symbol: end_symbol
})
_, cost, step = sess.run([ntm.optims[seq_length],
ntm.get_loss(seq_length),
ntm.global_step], feed_dict=feed_dict)
if idx % 100 == 0:
ntm.save(config.checkpoint_dir, 'copy', step)
if idx % print_interval == 0:
print("[%5d] %2d: %.2f (%.1fs)" \
% (idx, seq_length, cost, time.time() - start_time))
print("Training Copy task finished")
return cell, ntm
def __init__(self,
num_inputs,
num_outputs,
controller_size,
controller_layers,
num_heads,
N,
M,
controller_type='lstm'):
"""Initialize an EncapsulatedNTM.
:param num_inputs: External number of inputs.
:param num_outputs: External number of outputs.
:param controller_size: The size of the internal representation.
:param controller_layers: Controller number of layers.
:param num_heads: Number of heads.
:param N: Number of rows in the memory bank.
:param M: Number of cols/features in the memory bank.
"""
super(EncapsulatedNTM, self).__init__()
# Save args
self.num_inputs = num_inputs
self.num_outputs = num_outputs
self.controller_size = controller_size
self.controller_layers = controller_layers
self.num_heads = num_heads
self.N = N
self.M = M
# Create the NTM components
memory = NTMMemory(N, M)
if controller_type == 'lstm':
controller = LSTMController(num_inputs + M * num_heads,
controller_size, controller_layers)
else:
controller = MLPController(num_inputs + M * num_heads,
controller_size, controller_layers)
heads = nn.ModuleList([])
for i in range(num_heads):
heads += [
NTMReadHead(memory, controller_size),
NTMWriteHead(memory, controller_size)
]
self.ntm = NTM(num_inputs, num_outputs, controller, memory, heads)
self.memory = memory
def create_ntm(FLAGS, sess, **ntm_args):
cell = NTMCell(
input_dim=FLAGS.input_dim,
output_dim=FLAGS.output_dim,
controller_layer_size=FLAGS.controller_layer_size,
write_head_size=FLAGS.write_head_size,
read_head_size=FLAGS.read_head_size)
ntm = NTM(
cell, sess, FLAGS.min_length, FLAGS.max_length,
test_max_length=FLAGS.test_max_length, scope='NTM-%s' % FLAGS.task, **ntm_args)
return cell, ntm
def create_ntm(config, sess, **ntm_args):
cell = NTMCell(
input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
controller_dim=config.controller_dim,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size)
scope = ntm_args.pop('scope', 'NTM-%s' % config.task)
ntm = NTM(
cell, sess, config.min_length, config.max_length,
test_max_length=config.test_max_length, scope=scope, **ntm_args)
return cell, ntm
def compare():
obstacle, wall_cw, wall_awc = Obstacle(), WallCW(), WallACW()
obstacle_params, wall_cw_params, wall_acw_params = torch.load(
'program_memory/move.pt'), torch.load(
'program_memory/cw.pt'), torch.load('program_memory/acw.pt')
networks = [obstacle, wall_cw, wall_awc]
params = [obstacle_params, wall_cw_params, wall_acw_params]
hnm = HNM(10, 14, networks, params)
hnm.load_state_dict(torch.load('learned_params/hnm.pt'))
ntm = NTM(10, 14)
ntm.load_state_dict(torch.load('learned_params/ntm.pt'))
lstm = LSTM(14, 64, 3, 1)
lstm.load_state_dict(torch.load('learned_params/lstm.pt'))
testX, testY = getTestData()
hnm_correct, ntm_correct, lstm_correct = 0, 0, 0
totSamples = 0
for i in range(0, 25):
s = torch.from_numpy(np.array(testX[i:i + 1][0])).float().unsqueeze(0)
s_lstm = s.view(s.size()[0], s.size()[2], -1)
l = np.array(testY[i:i + 1][0])
print(i)
(hnm_read_weights, hnm_write_weights) = hnm._initialise()
(ntm_read_weights, ntm_write_weights) = ntm._initialise()
lstm_h = lstm.h0.expand(s_lstm.size()[0], 64)
lstm_c = lstm.c0.expand(s_lstm.size()[0], 64)
for j in range(s.size()[1]):
(hnm_out, hnm_read_weights,
hnm_write_weights) = hnm.forward(s[:, j, :], hnm_read_weights,
hnm_write_weights)
(ntm_out, ntm_read_weights,
ntm_write_weights) = ntm.forward(s[:, j, :], ntm_read_weights,
ntm_write_weights)
lstm_h, lstm_c, lstm_out = lstm.forward(s_lstm[:, :, j], lstm_h,
lstm_c)
if np.argmax(hnm_out.detach().numpy()) == np.argmax(l[j]):
hnm_correct += 1
if np.argmax(ntm_out.detach().numpy()) == np.argmax(l[j]):
ntm_correct += 1
if np.argmax(lstm_out.detach().numpy()) == np.argmax(l[j]):
lstm_correct += 1
totSamples += 1
print(hnm_correct, ntm_correct, lstm_correct)
print(totSamples)
def gen_model(input_dim,
batch_size,
output_dim,
n_slots=n_slots,
m_depth=m_depth,
controller_model=None,
activation="sigmoid",
read_heads=1,
write_heads=1):
model = Sequential()
model.name = "NTM_-_" + controller_model.name
model.batch_size = batch_size
model.input_dim = input_dim
model.output_dim = output_dim
ntm = NTM(output_dim,
n_slots=n_slots,
m_depth=m_depth,
shift_range=3,
controller_model=controller_model,
activation=activation,
read_heads=read_heads,
write_heads=write_heads,
return_sequences=True,
input_shape=(None, input_dim),
batch_size=batch_size)
model.add(ntm)
sgd = Adam(lr=learning_rate, clipnorm=clipnorm)
model.compile(loss='binary_crossentropy',
optimizer=sgd,
metrics=['binary_accuracy'],
sample_weight_mode="temporal")
return model
return parser.parse_args()
if __name__ == "__main__":
args = parse_arguments()
writer = SummaryWriter()
dataset = BinaySeqDataset(args)
dataloader = DataLoader(dataset, batch_size=1,
shuffle=True, num_workers=4)
model = NTM(M=args.memory_capacity,
N=args.memory_vector_size,
input_size=args.token_size,
output_size=args.token_size,
controller_out_dim=args.controller_output_dim,
controller_hid_dim=args.controller_hidden_dim,
)
print(model)
criterion = torch.nn.BCELoss()
optimizer = torch.optim.RMSprop(model.parameters(), lr=args.learning_rate)
print("--------- Number of parameters -----------")
print(model.calculate_num_params())
print("--------- Start training -----------")
losses = []
class NTMAugmentedDecoder(nn.Module):
def __init__(self,
d_vocab,
d_emb,
d_dec,
max_len,
bos_idx,
num_heads=8,
N=64,
M=32,
seg_size=20):
super().__init__()
self.d_vocab = d_vocab
self.seg_size = seg_size
self.embs = nn.Embedding(d_vocab, d_emb)
self.rnn = nn.GRU(d_emb, d_dec, batch_first=True)
self.ntm_scale = nn.Parameter(torch.zeros([1, d_dec]),
requires_grad=True)
self.ntm = NTM('mem-aug',
embedding_size=d_dec,
hidden_size=d_dec,
memory_size=M,
head_num=num_heads,
memory_feature_size=N,
output_size=d_dec)
self.init = nn.Parameter(torch.zeros(1, d_dec), requires_grad=True)
self.bos_idx = nn.Parameter(torch.tensor([bos_idx]),
requires_grad=False)
self.out_layer = nn.Linear(d_dec, d_vocab)
self.max_len = max_len
def forward(self, labels, state=None):
"""
Run decoder on input context with optional conditioning on labels and prelabels
:param labels: labels conditioned on at each timestep in next-prediction task (used during training)
:param state: initial state to begin decoding with. Could be output of encoder.
:return: If labels are provided, returns logits tensor (batch_size, num_steps, vocab_size). If labels are not provided,
returns predictions tensor (batch_size, num_steps) using provided sampling function.
"""
batch_size = labels.shape[0]
self.ntm.reset(batch_size, device=labels.device)
if state is None:
state = self.init.expand(batch_size, -1).contiguous() # bxh
init = self.embs(self.bos_idx.expand(batch_size).unsqueeze(1)) # bx1xh
# initialize ntm state, which keeps track of reads and writes
ntm_state = torch.zeros_like(state).to(state.device)
labels_no_last = labels[:, :
-1] # b x (t-1) # we don't take last word as input
# break input into slices, read and write between slices
num_slices = labels.shape[1] // self.seg_size
all_logit_slices = []
for slice in range(num_slices):
# grab slice of input
labels_slice = labels_no_last[:, slice *
self.seg_size:slice * self.seg_size +
self.seg_size]
in_embs = self.embs(labels_slice) # b x (t-1) x w
if slice == 0: # add bos index on first iteration
in_embs = torch.cat([init, in_embs], dim=1) # b x t x w
# give ntm state as input to all time steps of next slice
# multiple ntm state by scalars before giving to RNN, so it is not used in the beginning of training
#scaled_ntm_state = ntm_state * self.ntm_scale
#exp_ntm_state = scaled_ntm_state.unsqueeze(1).expand([-1, in_embs.shape[1], -1]) # b x t x h
rnn_input = in_embs # torch.cat([in_embs, exp_ntm_state], dim=-1)
# read slice of conversation history, with access to ntm state
outputs, _ = self.rnn(
rnn_input, state.unsqueeze(0)) # b x (t-1) x h OR b x t x h
# grab last state and use it to read and write from ntm
state = outputs[:, -1, :]
#ntm_state = self.ntm(state)
# predict outputs for this slice
logits = self.out_layer(outputs) # b x t x v
all_logit_slices.append(logits)
# append predictions for all slices together
logits = torch.cat(all_logit_slices, dim=1)
return logits
def complete(self, x, state=None, sample_func=None):
"""
Given tensor x containing token indices, fill in all padding token (zero) elements
with predictions from the NTM decoder.
:param x: (batch_size, num_steps) tensor containing token indices
:return: tensor same shape as x, where zeros have been filled with decoder predictions
"""
batch_size, num_steps = x.shape
self.ntm.reset(batch_size, device=x.device)
if state is None:
state = self.init.expand(batch_size, -1).contiguous() # bxh
if sample_func is None:
sample_func = partial(torch.argmax, dim=-1)
ntm_state = torch.zeros_like(state).to(state.device)
all_logits = []
all_preds = []
init = self.embs(self.bos_idx.expand(batch_size).unsqueeze(1)) # bx1xh
word = init.squeeze(1) # b x w
for step in range(num_steps):
# run RNN over input words
rnn_input = word.unsqueeze(
1) # torch.cat([word, ntm_state], dim=-1).unsqueeze(1)
_, state = self.rnn(rnn_input, state.unsqueeze(0)) # 1 x b x h
state = state.squeeze(0)
# produce prediction at each time step
logits = self.out_layer(state) # b x v
all_logits.append(logits)
pred = sample_func(logits) # b
# # at the end of each segment, read and write from NTM
# if step % self.seg_size == (self.seg_size - 1):
# # end of each segment, read and write from NTM
# ntm_state = self.ntm(state)
# here, we grab word from x if it exists, otherwise use prediction
mask = (x[:, step] != 0).long() # b
word_index = x[:, step] * mask + pred * (
1 - mask) # use label or prediction, whichever is available
word = self.embs(word_index) # b x w
all_preds.append(word_index)
logits = torch.stack(all_logits, dim=1)
return logits, torch.stack(all_preds, dim=1)
cur_dir = os.getcwd()
PATH = os.path.join(cur_dir, args.saved_model)
# PATH = os.path.join(cur_dir, 'saved_models/saved_model_copy_500000.pt')
# ntm = torch.load(PATH)
"""
For the Copy task, input_size: seq_width + 2, output_size: seq_width
For the RepeatCopy task, input_size: seq_width + 2, output_size: seq_width + 1
For the Associative task, input_size: seq_width + 2, output_size: seq_width
For the NGram task, input_size: 1, output_size: 1
For the Priority Sort task, input_size: seq_width + 1, output_size: seq_width
"""
ntm = NTM(input_size=task_params['seq_width'] + 1,
output_size=task_params['seq_width'],
controller_size=task_params['controller_size'],
memory_units=task_params['memory_units'],
memory_unit_size=task_params['memory_unit_size'],
num_heads=task_params['num_heads'])
ntm.load_state_dict(torch.load(PATH))
# -----------------------------------------------------------------------------
# --- evaluation
# -----------------------------------------------------------------------------
ntm.reset()
data = dataset[0] # 0 is a dummy index
input, target = data['input'], data['target']
out = torch.zeros(target.size())
# -----------------------------------------------------------------------------
# loop for other tasks
task_params['num_heads'], task_params['uniform'],
task_params['random_distr'], task_params['multi_layer_controller'])
# Output directory for tensorboard
configure(args.tb_dir + "/" + saved_model_name)
"""
For the Copy task, input_size: seq_width + 2, output_size: seq_width
For the RepeatCopy task, input_size: seq_width + 2, output_size: seq_width + 1
For the Associative task, input_size: seq_width + 2, output_size: seq_width
For the NGram task, input_size: 1, output_size: 1
For the Priority Sort task, input_size: seq_width + 1, output_size: seq_width
"""
ntm = NTM(input_size=task_params['seq_width'] + 1,
output_size=task_params['seq_width'],
controller_size=task_params['controller_size'],
memory_units=task_params['memory_units'],
memory_unit_size=task_params['memory_unit_size'],
num_heads=task_params['num_heads'],
multi_layer_controller=task_params['multi_layer_controller'])
if args.load_model != "":
ntm.load_state_dict(torch.load(args.load_model))
criterion = nn.BCELoss()
# As the learning rate is task specific, the argument can be moved to json file
optimizer = optim.RMSprop(ntm.parameters(),
lr=args.lr,
alpha=args.alpha,
momentum=args.momentum)
'''
optimizer = optim.Adam(ntm.parameters(), lr=args.lr,
controller_input_dim, controller_output_dim = controller_shape(num_encoder_tokens,
layer_dim,
m_depth,
n_slots,
shift_range,
read_heads,
write_heads)
encoder_inputs = Input(shape=(None, num_encoder_tokens))
encoder = NTM(layer_dim,
n_slots=n_slots,
m_depth=m_depth,
shift_range=shift_range,
controller_model=None,
activation="sigmoid",
read_heads = read_heads,
write_heads = write_heads,
return_sequences=True,
return_state=True)
saidas = encoder(encoder_inputs)
print(saidas[1])
def generate_target_original_plots(iteration, task_params, model_path,
image_output):
dataset = PrioritySort(task_params)
criterion = nn.BCELoss()
ntm = NTM(input_size=task_params['seq_width'] + 1,
output_size=task_params['seq_width'],
controller_size=task_params['controller_size'],
memory_units=task_params['memory_units'],
memory_unit_size=task_params['memory_unit_size'],
num_heads=task_params['num_heads'],
save_weigths=True,
multi_layer_controller=task_params['multi_layer_controller'])
ntm.load_state_dict(torch.load(model_path))
# -----------------------------------------------------------------------------
# --- evaluation
# -----------------------------------------------------------------------------
ntm.reset()
data = dataset[0] # 0 is a dummy index
input, target = data['input'], data['target']
out = torch.zeros(target.size())
# -----------------------------------------------------------------------------
# loop for other tasks
# -----------------------------------------------------------------------------
for i in range(input.size()[0]):
# to maintain consistency in dimensions as torch.cat was throwing error
in_data = torch.unsqueeze(input[i], 0)
ntm(in_data)
# passing zero vector as the input while generating target sequence
in_data = torch.unsqueeze(torch.zeros(input.size()[1]), 0)
for i in range(target.size()[0]):
out[i] = ntm(in_data)
loss = criterion(out, target)
binary_output = out.clone()
binary_output = binary_output.detach().apply_(lambda x: 0
if x < 0.5 else 1)
# sequence prediction error is calculted in bits per sequence
error = torch.sum(torch.abs(binary_output - target))
fig = plt.figure()
ax1 = fig.add_subplot(211)
ax2 = fig.add_subplot(221)
ax1.set_title("Result")
ax2.set_title("Target")
sns.heatmap(binary_output,
ax=ax1,
vmin=0,
vmax=1,
linewidths=.5,
cbar=False,
square=True)
sns.heatmap(target,
ax=ax2,
vmin=0,
vmax=1,
linewidths=.5,
cbar=False,
square=True)
plt.savefig(
image_output +
"/priority_sort_{}_{}_{}_{}_{}_{}_{}_{}_{}_image_{}.png".format(
task_params['seq_width'] + 1, task_params['seq_width'],
task_params['controller_size'], task_params['memory_units'],
task_params['memory_unit_size'], task_params['num_heads'],
task_params['uniform'], task_params['random_distr'],
task_params['multi_layer_controller'], iteration))
fig = plt.figure(figsize=(15, 6))
ax1_2 = fig.add_subplot(211)
ax2_2 = fig.add_subplot(212)
ax1_2.set_title("Read Weigths")
ax2_2.set_title("Write Weights")
sns.heatmap(ntm.all_read_w, ax=ax1_2, linewidths=.01, square=True)
sns.heatmap(ntm.all_write_w, ax=ax2_2, linewidths=.01, square=True)
plt.tight_layout()
plt.savefig(
image_output +
"/priority_sort_{}_{}_{}_{}_{}_{}_{}_{}_{}_weigths_{}.png".format(
task_params['seq_width'] + 1, task_params['seq_width'],
task_params['controller_size'], task_params['memory_units'],
task_params['memory_unit_size'], task_params['num_heads'],
task_params['uniform'], task_params['random_distr'],
task_params['multi_layer_controller'], iteration),
dpi=250)
# ---logging---
print('[*] Checkpoint Loss: %.2f\tError in bits per sequence: %.2f' %
(loss, error))
def get_ntm_model():
from keras.models import Model
import keras.backend as K
assert permute_layer is not None
num_read = 1
num_write = 1
mem_length = 40
n_slots = 128
model_input = Input(
(WINDOW_LENGTH, 1) + INPUT_SHAPE,
#batch_shape = (batch_size,) + (WINDOW_LENGTH,1) + INPUT_SHAPE
)
per = permute_layer(model_input)
x = TimeDistributed(
Conv2D(32, (8, 8), name='conv1', activation='relu',
subsample=(4, 4)))(per)
x = TimeDistributed(
Conv2D(64, (4, 4), name='conv2', activation='relu',
subsample=(2, 2)))(x)
x = TimeDistributed(
Conv2D(64, (3, 3), name='conv3', activation='relu',
subsample=(1, 1)))(x)
#x = TimeDistributed(Conv2D(64,(4,4),name='conv2',activation = 'relu',subsample = (2,2)))(x)
#x = TimeDistributed(Conv2D(64,(3,3),name='conv3',activation = 'relu',subsample = (1,1)))(x)
x = TimeDistributed(Flatten(name="Flatten1"))(
x) # (batch_size,WINDOW_LENGTH,3176)
x_shape = K.int_shape(x)
print('x has shape:', x_shape)
# controller construction
controller_inp = Input((x_shape[-1], ),
name="controller_input") # (batch_size,3176)
read_inp = Input((num_read, mem_length),
name="read_inp") # (batch_size,n_read,n_write)
read_inp_flatten = Flatten(name="read_inp_flatten")(
read_inp) #(batch_size,n_read * n_write)
print('controller_inp shape:', controller_inp.shape)
#print('read_inp_flatten shape:',K.int_shape(read_inp_flatten))
#print('read_inp_flatten_repeat shape:',K.int_shape(read_inp_flatten_repeat))
#hidden_int = Dense(512,activation = 'relu')(controller_inp)
hidden = Concatenate(name="ctrl_inp_read_inp_concat")(
[controller_inp,
read_inp_flatten]) # (batch_size, 3176 + num_read * mem_length)
#hidden = Dense(512,activation = 'relu')(concat)
controller_output = Dense(nb_actions, activation='linear')(hidden)
controller = Model([controller_inp, read_inp],
[controller_output, controller_inp])
controller.summary()
# ntm constuction
#TODO: reset the state for on_batch_end!!
ntm_cell = NTM(
controller, # custom controller, should output a vector
n_slots,
mem_length, # Memory config
num_shift=3, # shifting
batch_size=batch_size,
#controller_instr_output_dim = controller_instr_output_dim,
return_sequences=False,
is_controller_recurrent=True,
num_read=num_read,
num_write=num_write)(x) # (batch_size,512)
ntm_cell_output_shape = K.int_shape(ntm_cell)
print('ntm_cell output:', ntm_cell_output_shape)
ntm_cell_output_shape = ntm_cell_output_shape[1:]
#model_output = Dense(nb_actions,activation = 'linear')(ntm_cell)
model = Model(model_input, ntm_cell)
model.summary()
return model
n_slots = n_slots,
m_depth = m_depth,
controller_model = None,
activation = "sigmoid",
read_head = 1,
write_head = 1):
model = Sequential()
model.name = "NTM_-_"+ controller_model.name
model.batch_size = batch_size
model.input_dim = input_dim
model.ouput_dim =output_dim
ntm = NTM(output_dim, n_slots = n_slots, m_depth = m_depth,
shift_range = 3,
controller_model = controller_model,
activation = activation,
read_heads = read_heads,
write_heads = write_heads,
# return_sequences = True,
input_shape = (None,input_dim),
batch_size = batch_size)
model.add(NTM)
sgd = Adam(lr = learning_rate, clipnorm = clipnorm)
model.compile(loss = 'binary_crossentropy',optimizer=sgd, metrics = ['binary_accuracy'],sample_weight_model = "temporal")
return model
from keras.models import Sequential
from keras.optimizers import Adam
from ntm import NeuralTuringMachine as NTM
model = Sequential()
ntm = NTM([625],
n_slots=50,
m_depth=20,
shift_range=3,
controller_model=None,
return_sequences=True,
input_shape=(None, 625),
batch_size=100)
model.add(ntm)
# sgd = Adam(lr=learning_rate, clipnorm=clipnorm)
model.compile(loss='binary_crossentropy',
optimizer='Adam',
metrics=['binary_accuracy'],
sample_weight_mode="temporal")
print(model.summary())
def trainNTM():
ntm = NTM(10, 14)
X, y = getData()
ntm.train(X, y, 1)
dataset = PrioritySort(task_params)
input_size=task_params['seq_width']+1
output_size=task_params['seq_width']
"""
For the Copy task, input_size: seq_width + 2, output_size: seq_width
For the RepeatCopy task, input_size: seq_width + 2, output_size: seq_width + 1
For the Associative task, input_size: seq_width + 2, output_size: seq_width
For the NGram task, input_size: 1, output_size: 1
For the Priority Sort task, input_size: seq_width + 1, output_size: seq_width
"""
has_tau=0
if args.model=='ntm':
model = NTM(input_size= input_size,
output_size=output_size,
controller_size=args.lstm_size,
memory_units=128,
memory_unit_size=20,
num_heads=1)#task_params['num_heads'])
elif args.model=='dnc':
model = DNC(input_size= input_size,
output_size=output_size,
hidden_size=args.lstm_size,
nr_cells=128,
cell_size=20,
read_heads=1)#task_params['num_heads'])
model.init_param()
elif args.model=='sam':
model = SAM(input_size= input_size,
output_size=output_size,
hidden_size=args.lstm_size,
nr_cells=128,
'length': 5,
'controller_layer_size': 1,
'write_head_size': 1,
'read_head_size': 1,
'checkpoint_dir': 'checkpoint'
}
if __name__ == "__main__":
with tf.device('/cpu:0'), tf.Session() as sess:
cell = NTMCell(input_dim=config['input_dim'],
output_dim=config['output_dim'],
controller_layer_size=config['controller_layer_size'],
write_head_size=config['write_head_size'],
read_head_size=config['read_head_size'],
controller_dim=32)
ntm = NTM(cell, sess, config['length'] * 2 + 2)
if not os.path.isdir(config['checkpoint_dir'] + '/copy_' +
str(config['length'] * 2 + 2)):
print(" [*] Initialize all variables")
tf.global_variables_initializer().run()
print(" [*] Initialization finished")
else:
ntm.load(config['checkpoint_dir'], 'copy')
start_time = time.time()
print('')
for idx in range(config['epoch']):
seq_length = np.random.randint(2, config['length'] + 1)
X, Y, masks = build_seq_batch(seq_length, config['length'],
config['input_dim'] - 2)
class EncapsulatedNTM(nn.Module):
def __init__(self,
num_inputs,
num_outputs,
controller_size,
controller_layers,
num_heads,
N,
M,
controller_type='lstm'):
"""Initialize an EncapsulatedNTM.
:param num_inputs: External number of inputs.
:param num_outputs: External number of outputs.
:param controller_size: The size of the internal representation.
:param controller_layers: Controller number of layers.
:param num_heads: Number of heads.
:param N: Number of rows in the memory bank.
:param M: Number of cols/features in the memory bank.
"""
super(EncapsulatedNTM, self).__init__()
# Save args
self.num_inputs = num_inputs
self.num_outputs = num_outputs
self.controller_size = controller_size
self.controller_layers = controller_layers
self.num_heads = num_heads
self.N = N
self.M = M
# Create the NTM components
memory = NTMMemory(N, M)
if controller_type == 'lstm':
controller = LSTMController(num_inputs + M * num_heads,
controller_size, controller_layers)
else:
controller = MLPController(num_inputs + M * num_heads,
controller_size, controller_layers)
heads = nn.ModuleList([])
for i in range(num_heads):
heads += [
NTMReadHead(memory, controller_size),
NTMWriteHead(memory, controller_size)
]
self.ntm = NTM(num_inputs, num_outputs, controller, memory, heads)
self.memory = memory
def init_sequence(self, batch_size):
"""Initializing the state."""
self.batch_size = batch_size
self.memory.reset(batch_size)
self.previous_state = self.ntm.create_new_state(batch_size)
def forward(self, x=None):
if x is None:
x = Variable(torch.zeros(self.batch_size, self.num_inputs))
o, self.previous_state = self.ntm(x, self.previous_state)
return o, self.previous_state
def calculate_num_params(self):
"""Returns the total number of parameters."""
num_params = 0
for p in self.parameters():
num_params += p.data.view(-1).size(0)
return num_params
dataset = RepeatCopyDataset(task_params)
dataset = AssociativeDataset(task_params)
dataset = NGram(task_params)
dataset = PrioritySort(task_params)
'''
"""
For the Copy task, input_size: seq_width + 2, output_size: seq_width
For the RepeatCopy task, input_size: seq_width + 2, output_size: seq_width + 1
For the Associative task, input_size: seq_width + 2, output_size: seq_width
For the NGram task, input_size: 1, output_size: 1
For the Priority Sort task, input_size: seq_width + 1, output_size: seq_width
"""
ntm = NTM(input_size=task_params['seq_width'] + 2,
output_size=task_params['seq_width'],
controller_size=task_params['controller_size'],
memory_units=task_params['memory_units'],
memory_unit_size=task_params['memory_unit_size'],
num_heads=task_params['num_heads'])
criterion = nn.BCELoss()
# As the learning rate is task specific, the argument can be moved to json file
optimizer = optim.RMSprop(ntm.parameters(),
lr=args.lr,
alpha=args.alpha,
momentum=args.momentum)
'''
optimizer = optim.Adam(ntm.parameters(), lr=args.lr,
betas=(args.beta1, args.beta2))
'''
args.saved_model = 'saved_model_copy.pt'
task_params['max_seq_len'] = data[d][1]
dataset = ReverseDataset(task_params)
dataset2 = ReverseDataset(task_params)
task_params['min_seq_len'] = data[d][2]
task_params['max_seq_len'] = data[d][3]
dataset3 = ReverseDataset(task_params)
#4.save model
args.saved_model = 'saved_model/' + 'saved_model_reverse_' + args.config + '.pt'
cur_dir = os.getcwd()
PATH = os.path.join(cur_dir, args.saved_model)
"""
For the Reverse task, input_size: seq_width + 2, output_size: seq_width
"""
ntm = NTM(input_size=task_params['seq_width'] + 2,
output_size=task_params['seq_width'],
controller_size=task_params['controller_size'],
memory_units=task_params['memory_units'],
memory_unit_size=task_params['memory_unit_size'],
num_heads=task_params['num_heads'])
criterion = nn.BCELoss()
# As the learning rate is task specific, the argument can be moved to json file
# optimizer = optim.RMSprop(ntm.parameters(),
# lr=args.lr,
# alpha=args.alpha,
# momentum=args.momentum)
optimizer = optim.Adam(ntm.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))
# ----------------------------------------------------------------------------
# -- basic training loop
model = Sequential()
model.name = "NTM_-_" + None.name
model.batch_size = batch_size
model.input_dim = input_dim
model.output_dim = 1
ntm = NTM(
1,
n_slots=n_slots, #n_slots: Memory width
m_depth=m_depth, #m_depth: Memory depth at each location
shift_range=3,
#shift_range: int, number of available shifts, ex. if 3, avilable shifts are
# (-1, 0, 1)
controller_model=None,
#controller_model: A keras model with required restrictions to be used as a controller.
# The requirements are appropriate shape, linear activation and stateful=True if recurrent.
# Default: One dense layer.
activation="sigmoid",
#activation: This is the activation applied to the layer output.
# It can be either a Keras activation or a string like "tanh", "sigmoid", "linear" etc.
# Default is linear.
read_heads=1,
write_heads=1,
return_sequences=True,
input_shape=(None, input_dim),
batch_size=batch_size)
model.add(ntm)
model.compile(loss='binary_crossentropy',
optimizer=Adam(lr=learning_rate, clipnorm=clipnorm),
metrics=['binary_accuracy'],
sample_weight_mode="temporal")
def main():
train = True
weight_path = "" #"model-constant-mem_4500.save"
if train:
# The big cheese, we're doin it guys!
plt.ion()
fig = plt.figure()
# COPY TASK, copy a 20 length tensor of random numbers. Each section will contain 5 bits, one of which (bits 0 - 3) will be on.
# We'll input the 20 length, then input an end token, which will be the 5th bit (or 4th if using zero-based indexing), then input another 20 length with all zeros. This signals to the NTM that it needs to start outputting the copy.
class Controller:
def __init__(self):
# Output size is 5, because it needs to output the copied 5 bits
self.size = 128
# We'll have 1 read head, which produces a single read_vector of size 10. We also need to feed in the input, which is of size 5 (for the five bits)
# so our total input size is 15
self.fc_1 = init_weight(15, 128)
self.fc_2 = init_weight(128, 128)
# This is our controller output
self.fc_3 = init_weight(128, 128)
def get_weights(self):
return [self.fc_1, self.fc_2, self.fc_3]
def forward(self, inp):
fc1 = T.nnet.relu(T.dot(inp, self.fc_1))
fc2 = T.nnet.relu(T.dot(fc1, self.fc_2))
# I would ReLU the output, but I already did in the NTM implementation
fc3 = T.dot(fc2, self.fc_3)
return fc3
# output size is 5, for the 5 copy bits
ntm = NTM(controller=Controller(),
output_size=5,
memory_slots=20,
slot_size=10,
read_heads=1,
batch_size=10)
data = T.tensor3()
target = T.tensor3()
#r = theano.shared(np.random.randn(10, 10, 20))
r = theano.shared(1.)
r_ = theano.shared(np.zeros([10, 10, 20])) + r
if weight_path != '':
print("loading weights")
# Load weights, but just the NTM weights, not memory, since we may extend it
checkpoint = open(weight_path, 'rb')
all_weights = ntm.weights
for w in all_weights:
w.set_value(cPickle.load(checkpoint).get_value())
checkpoint.close()
memory_states, _, weightings, ntm_outputs = ntm.process(data, r_)
# We average the loss across batches, so that we have a singular loss for each timestep. We then average these losses
# ntm_outputs - target ** 2 -> ts x batchsize x bits
#
loss = T.sum(T.mean(T.sum(5 *
(T.nnet.sigmoid(ntm_outputs) - target)**2,
axis=2),
axis=1),
axis=0)
updates = RMSprop(cost=loss, params=ntm.weights + [r], lr=1e-3)
train = theano.function(inputs=[data, target],
outputs=[
memory_states, weightings, weightings,
ntm_outputs, loss, updates[2][1]
],
updates=updates)
for example in range(5000):
# Produce the first half
# let's feed a test example
# ts x batchsize x bits
end = np.zeros([1, 10, 5])
for batch in range(10):
end[0, batch, -1] = 1 # Make the last bit in each batch a 1
first_half = (np.random.randn(10, 10, 5) > 0).astype(
np.float32) * 1
for batch in range(10):
first_half[:, batch,
-1] = 0 # Make sure the last bit (end bit) of each batch is 0
# Produce second half
second_half = np.zeros([10, 10, 5]) # Just a bunch of zeros
data = np.concatenate([first_half, end, second_half], axis=0)
target = np.concatenate([second_half, end, first_half], axis=0)
# lamar gotta have that extra timestep for the end bit
outputs = train(data, target)
print("LOSS " + str(outputs[-2]) + ", " + str(example))
read = outputs[2]
read = read[:, 0, 0, :]
write = outputs[2]
write = write[:, 1, 0, :]
outputs = outputs[3]
outputs = outputs[:, 0]
#.transpose([1, 0])
if (example % 20 == 0 and example != 0):
cmap = 'jet'
fig.add_subplot(2, 2, 1)
plt.imshow(sigmoid(outputs), cmap=cmap)
fig.add_subplot(2, 2, 2)
plt.imshow(target[:, 0], cmap=cmap)
fig.add_subplot(2, 2, 3)
plt.imshow(read, cmap=cmap)
fig.add_subplot(2, 2, 4)
plt.imshow(write, cmap=cmap)
plt.pause(0.1)
if (example % 500 == 0):
print("SAVING WEIGHTS")
f = open('model-constant-mem_' + str(example) + '.save', 'wb')
for w in ntm.weights + [r]:
cPickle.dump(w, f, protocol=cPickle.HIGHEST_PROTOCOL)
f.close()
"""
fig = plt.figure()
fig.add_subplot(2, 2, 1)
plt.imshow(data[:, 0, :], origin = [0, 0])
fig.add_subplot(2, 2, 2)
plt.imshow(target[:, 0, :], origin = [10, 0])
plt.show()
"""
else:
# Test time!
plt.ion()
fig = plt.figure()
# COPY TASK, we're going to see how well our Neural Turing Machine model extends to longer sequences
class Controller:
def __init__(self):
# Output size is 5, because it needs to output the copied 5 bits
self.size = 128
# We'll have 1 read head, which produces a single read_vector of size 10. We also need to feed in the input, which is of size 5 (for the five bits)
# so our total input size is 15
self.fc_1 = init_weight(15, 128)
self.fc_2 = init_weight(128, 128)
# This is our controller output
self.fc_3 = init_weight(128, 128)
def get_weights(self):
return [self.fc_1, self.fc_2, self.fc_3]
def forward(self, inp):
fc1 = T.nnet.relu(T.dot(inp, self.fc_1))
fc2 = T.nnet.relu(T.dot(fc1, self.fc_2))
# I would ReLU the output, but I already did in the NTM implementation
fc3 = T.dot(fc2, self.fc_3)
return fc3
# output size is 5, for the 5 copy bits
ntm = NTM(controller=Controller(),
output_size=5,
memory_slots=80,
slot_size=10,
read_heads=1,
batch_size=10)
data = T.tensor3()
# Load weights
checkpoint = open(
'pretrained-models-copy/model-constant-mem_4500.save', 'rb')
all_weights = ntm.weights
for w in all_weights:
w.set_value(cPickle.load(checkpoint).get_value())
r = theano.shared(
np.zeros(shape=[10, 10, 80]) +
cPickle.load(checkpoint).get_value())
checkpoint.close()
memory_states, _, weightings, ntm_outputs = ntm.process(data, r)
test = theano.function(
inputs=[data],
outputs=[memory_states, weightings, weightings, ntm_outputs])
for example in range(5000):
print(r.get_value())
# Produce the first half
# let's feed a test example
# ts x batchsize x bits
end = np.zeros([1, 10, 5])
for batch in range(10):
end[0, batch, -1] = 1 # Make the last bit in each batch a 1
first_half = (np.random.randn(60, 10, 5) > .7).astype(
np.float32) * 1
for batch in range(10):
first_half[:, batch,
-1] = 0 # Make sure the last bit (end bit) of each batch is 0
# Produce second half
second_half = np.zeros([60, 10, 5]) # Just a bunch of zeros
data = np.concatenate([first_half, end, second_half], axis=0)
# lamar gotta have that extra timestep for the end bit
outputs = test(data)
read = outputs[2]
read = read[:, 0, 0, :]
write = outputs[2]
write = write[:, 1, 0, :]
outputs = outputs[3]
outputs = outputs[:, 0]
#.transpose([1, 0])
cmap = 'jet'
fig.add_subplot(2, 2, 1)
plt.imshow(sigmoid(outputs), cmap=cmap)
fig.add_subplot(2, 2, 2)
plt.imshow(data[:, 0], cmap=cmap)
fig.add_subplot(2, 2, 3)
plt.imshow(read, cmap=cmap)
fig.add_subplot(2, 2, 4)
plt.imshow(write, cmap=cmap)
plt.pause(0.1)
input("")
from_checkpoint = opt[1]
elif opt[0] == '--iterations':
iterations = int(opt[1])
graph = tf.Graph()
with graph.as_default():
with tf.compat.v1.Session(graph=graph) as session:
llprint("Building Computational Graph ... ")
optimizer = tf.compat.v1.train.RMSPropOptimizer(learning_rate,
momentum=momentum)
turing_machine = NTM(RecurrentController, input_size, output_size,
memory_size, word_size, read_heads,
shift_range, batch_size)
# squash the DNC output between 0 and 1
output, _ = turing_machine.get_outputs()
squashed_output = tf.clip_by_value(tf.sigmoid(output), 1e-6,
1. - 1e-6)
loss = binary_cross_entropy(squashed_output,
turing_machine.target_output)
summaries = []
gradients = optimizer.compute_gradients(loss)
for i, (grad, var) in enumerate(gradients):
if grad is not None:
summaries.append(